Claude Prompts: How to Get the Best From Anthropic Models

Claude models reward structure more than almost any other model family. A few Claude-specific habits noticeably improve output quality.

Use XML-style tags to structure prompts

Claude is specifically trained to respect XML-style delimiters. Wrap distinct parts of your prompt in named tags:

<instructions>Summarize the report.</instructions> <report>…</report> <format>Three bullet points, each under 20 words.</format>

Tags eliminate the classic failure where instructions bleed into content. They also make templates trivially maintainable.

Put long documents first, questions last

For long-context work, Claude performs best when the document comes first and the question or instruction comes at the end of the prompt. For very long inputs, also ask Claude to quote relevant passages before answering; it grounds the response and reduces hallucination.

Give Claude room to think

For analytical tasks, instruct Claude to reason step by step inside a dedicated tag before giving a final answer. Separating thinking from the answer produces more careful work and lets you hide the reasoning in production interfaces.

Prefill the response format

You can start Claude’s answer for it. Ending your prompt with the beginning of the desired output, such as an opening brace for JSON, strongly steers the completion format.

Watch prompt complexity

Claude handles long instructions well, but every model skips steps when instruction density gets extreme. If your prompt scores above 80 on a complexity score, split it into a chain. And when copying prompts between tools, scan for hidden characters with an invisible character detector; zero-width characters inside XML tags are a uniquely frustrating bug.

Claude vs ChatGPT prompting

The fundamentals transfer, but the emphasis differs: Claude responds strongly to XML structure and document-first ordering, while ChatGPT leans harder on system-message persona. Write structured prompts and both families reward you.